GB2290635A

GB2290635A - Active control of noise and vibration

Info

Publication number: GB2290635A
Application number: GB9511548A
Authority: GB
Inventors: Steve C Southward; Dino J Rosetti; Mark R Jolly
Original assignee: Lord Corp
Current assignee: Lord Corp
Priority date: 1994-06-18
Filing date: 1995-06-07
Publication date: 1996-01-03
Anticipated expiration: 2015-06-07
Also published as: GB2290635B; GB9511548D0; US5627896A

Abstract

An active system for controlling vibration or sound limits the output gain G from a gradient descent algorithm, such as an LMS algorithm, according to the relationship <IMAGE> and reduces one of first and second filter weights W1 and W0 such that G </= Gmax ensuring the output cancelling signal is never saturated. This eliminates the square wave shape which may impart unwanted harmonics or resonances to the dynamic system. In another aspect, the system eliminates drift of the output device toward saturation when situations are encountered such as singularities in the quadratic performance surface. This is accomplished by applying leakage factor r(k) to a function f(k) at or near a predetermined limit to avoid saturation of the output device. The function f(k) can be the gain G or the weights W(k). Another aspect is to provide a constant, optimum, and stable adaptation rate in an LMS system by providing an optimum adaptation coefficient mu opt that provides a constant, uniform, and optimum adaptation rate on a per period basis for all frequencies. Various combinations of the above are described and are useful in active control systems. <IMAGE>

Description

ACTIVE CONTROL OF NOISE AND VIBRATION Field of the Invention This invention relates to the area of active control of noise and vibration. Specifically, the invention relates to improvements in implementations of least mean square (LMS) and Filtered-x LMS algorithms for controlling active systems.

Badkgnaund of the Invention Active vibration isolation and sound isolation systems are known in the art. These systems utilize microprocessors (processors) to control or minimize mechanical vibration or ambient noise levels at a defined location or locations. Typical examples include noise or vibration experienced in an aircraft cabin or within an automobile passenger compartment. Generally, these active systems are responsive to at least one external input signal such as a synchronizing tachometer signal and/or error signal as supplied by various types of sensors such as microphones, accelerometers, etc.

Generally, these systems strive to reduce to zero or at least minimize the recurring sound and/or vibration.

Examples of such systems are taught in US Pat. Nos. 4,677,676 to Eriksson, 4,153,815 to Chaplin et al., 4,122,303 to Chaplin et al., 4,417,098 to Chaplin et al., 4,232,381 to Rennick et al., 4,083,433 to Geohegan, Jr. et al., 4,878,188 to Ziegler, Jr., 4,562,589 to Warnaka et al., 4,473,906 to Warnaka et al., 5,170,433 to Elliott et al., 4,689,821 to Salikuddin et al., 5,174,552 to Hodgson, and 5,133,527 to Chen et al., the disclosures of each are hereby incorporated by reference herein. In these types of systems, the algorithm used for control can be least mean square (LMS), Filtered-x LMS, or the like.

In active control systems of the above-mentioned type, it is generally required to have an input signal(s) which is supplied to an adaptive filter and/or a processor which is indicative of the frequency content and/or amplitude/phase of the input source, i.e., indicative of what is causing the disturbance. Particularly, for tonal control systems it is usually required to have two or more analog or digital waveforms, such as a sine and cosine wave, that are synchronized with, i.e., at the same frequency as, the input source signal for providing the appropriate information to the processor and/or adaptive filter.

A block diagram of a typical active control system is detailed in Fig. 1. The active control system 20 is comprised of an input source sensor 22 from which is derived a signal indicative of the source of the disturbance to be canceled, whether that disturbance is a mechanical vibration, a rotating unbalance, a source of acoustic noise, or the like. The sensor signal is conditioned by conditioner 24, such as by bandpass filtering and amplifying such that a usable signal is achieved. Conditioning of the signal may not be required as dictated by the application. The resulting signal is passed into a waveform generator 25. The waveform generator 25 provides the appropriate synchronized waveforms to the processor 28 and the adaptive filter 26.

For tonal control or control of a single vibrational input, the waveform generator 25 generates at least two analog or digital waves, such as a sine and cosine wave, which are synchronized in frequency with the input signal. The output from the waveform generator 25 is generally input to an adaptive filter 26 and generally also fed to a processor 28 to trigger the processor to perform the appropriate calculations. The input to the filter 26 generally has the form: xk = Asin((okT)+BCos((okT) where xk = input signal T = Sample period A = Amplitude of the real portion B = Amplitude of the imaginary portion The filter 26 output contains each of the sine and cosine components, adjusted in amplitude by the adaptive weights Wk.In a single input-single output (SISO) system, there is a single input xk and weights Wo and Wi are used to adaptively update the adaptive filter 26 in response to the error and/or input signal. The error sensor 30 provides a signal to the conditioner 24' which is indicative of the residual disturbance, i.e., the residual noise or vibration at the location of interest.

After being conditioned, the signal is inputted to the processor 28 where various calculations take place. The calculations take place according to an algorithm in order to produce updated adaptation coefficient weights. These updates to the weights occur according to an LMS algorithm, Filtered-x LMS algorithm, or the like. The output from the filter 26 is again conditioned by conditioner 24" and inputted to output device 32 to produce a output disturbance Yk at the location of interest. The magnitude of the output disturbance Yk in the ideal case is sufficient to minimize the mean square error ek2 at the sensor location. In other words, ideally, the error will be driven to zero.

Multiple input-multiple output (MIMO) systems are also known which utilize multiple input sensors, output devices, and error sensors.

One such system is described in US Pat No. 5,216,722 to Popovich, the disclosure of which is hereby incorporated by reference herein. MIMO systems function essentially in the same fashion as SISO systems, albeit they are more computationally intensive.

In classical least mean square (LMS) control as illustrated by the block diagram of Fig. 2, an error signal ek is generated as the difference between the desired dk and the actual filter output yk. In the ideal case, the error signal is zero and the filter output Yk is equal and opposite to the desired dk, thus completely canceling the disturbance (the vibration or sound pressure) at the predetermined location of interest. In classical LMS control, the error(s) ek and the input(s) Xk are used to adaptively update the weight vector Wk in a Finite Impulse Response (FIR) Filter according to the LMS algorithm.The weights are updated according to the relationship: Wk+1 = Wk - ekXk where Wk = adaptive filter weight ek = error signal Xk = the input signal = the adaptation coefficient which regulates the speed and stability of adaptation The error is derived from the equation ek=dk-yk Therefore, the error becomes ek=dk - XkTWk where Xk = transpose of the Xk vector dk= desired disturbance ek= error signal In most practical instances, it is desired to minimize the mean squared error (MSE) signal, i.e., the average power of the error signal.

Therefore, the performance function to be minimized becomes: ek2 = (dk + XkTW)2= dk2+WTXkXkTW - 2dkXkTW This equation results in a multi-dimensional plot or quadratic performance surface which generally has one or more minima, which will be illustrated later. The classical LMS algorithm is fully described by Bernard Widrow and Samuel D. Stearns, Adoptive Sisal Processing, Prentice-Hall, 1985.

The Filtered-x algorithm is a modification of the basic feedforward LMS algorithm. This modification is required wher dynamics exist between the adaptive filter output and the error sensor signals. The modification to the update equation includes an estimate of the plant transfer function P. This estimate of the plant P enables the appropriate approximation of the gradient of the performance or error surface.

The block diagram of Fig. 3 illustrates a typical Filtered-x LMS algorithm. The major difference, as compared to classical LMS, is inclusion of plant dynamics. The plant estimate P which is obtained through a system identification (ID) step, is obtained either at startup or in an on-line fashion. The update equation for Filtered-x is given by: W*+, = Wk - p e,R" where Rk = a filtered version of the input signal Typical implementations of the Filtered-x LMS algorithm design the adaptation coefficient 11 to be stable in the worst case. This is achieved by choosing a fixed or constant value of 11 which is typically very small.For a fixed value of 1l, and when the plant gain is very low, the adaptation rate will also be very slow, but still stable. When the plant gain is very low, the adaptation rate was also very slow, but still stable. When the plant gain is very high, such as at a resonance peak, the adaptation rate will be correspondingly high. Fig. 3A illustrates this phenomena. Typically, the adaptation coefficient KL is frequency insensitive, i.e. it is a constant of the proper magnitude that insures stability of convergence at the worst case condition.

This illustrates one of the shortcomings of classical LMS and Filtered-x LMS algorithms. The adaptation coefficient 11 is frequency insensitive. Thus, the adaptation rate must vary as a function of the frequency of the input disturbance because the plant transfer function P varies as a function of frequency. As a result of this, optimum convergence can never be achieved in these systems. US Pat. No. 4,931,977 to Kiemes describes a method of adaptive filtering that allows convergence in a fixed, predetermined amount of time by a Newton-method descent method, the disclosure of which is hereby incorporated by reference. Other references describing various adaptation methods are US Pat.Nos. 5,260,896 to Iwasaki, 5,243,624 to Paik et al., 5,136,531 to McCaslin, 5,124,626 to Thoen, 5,058,047 to Chung, 4,989,170 to Batruni et al., 4,939,685 to Feintuch, and 4,878,188 to Ziegler, Jr.

Furthermore, under certain conditions when utilizing LMS control, the output from the digital processor to the adaptive filter can exceed the saturation limits. For example, when large disturbances are encountered. Because the filters and processors are only capable of outputting signals within predetermined operating limits or voltages, this can result in a detrimental square wave or clipped form to the filter output.

Furthermore, when the output device is an actuator, operating the output device in this region may cause the actuator to overheat. Therefore, a solution to the clipping and overheating problems is needed.

Fig. 3B illustrates a typical saturated tonal waveform output. The output waveform is clipped, having a shape which approximates a square wave. Inputting this waveform into the actuator/output device can result in excitation of resonances or harmonics in the dynamic system. This is because the shape of the waveform will act as an impulsive input to the system, i.e., very similar to a hammer strike input which has a tendency to excite many resonant frequencies or modes. Furthermore, simply operating the output device in its nonlinear region may excite multiple harmonics. US Pat.Nos. 4,736,460 to Rilling, 5,259,033 to Goodings et al., 5,216,692 to Ling, 4,268,829 to Baurle et al. all describe various methods of controlling signal amplitude and signal power levels Another problem with these active control systems utilizing LMS control is that the performance surface can contain singularities.

Singularities can create serious performance problems when adaptive feedforward algorithms such as Filtered-x LMS are used in MIMO systems. It is generally assumed that the quadratic performance surface, upon which the gradient based algorithms are based, has a global minimum. However, when there is no one local minima on the performance surface, a singularity exists.

Fig. 4A illustrates such a situation where there are multiple local minima or a trough in the quadratic performance surface. Fig. 4B illustrates a typical system with only one minima. When there is a global minima, the control system will be easily converged to minimize the mean square error (MSE) using LMS, Filtered-x LMS, or the like. However, where there is no one unique minima, but several minima, or a trough as shown in Fig. 4A, the system output may become saturated. Because of inaccuracies in the plant estimates which affect the gradient estimate, as well as certain kinds of noise, the output device signals uk may drift toward the saturation limits when a singularity condition is encountered.

Therefore, this situation will impart unwanted harmonics into the system.

These harmonics are caused by signal clipping and from operating within the nonlinear working region of the output device. Therefore, there is a need for a method of compensating for singularity conditions. One method of preventing overdrive of a transducer is provided in US Pat. No. 5,278,913 to Delfosse et al.

Summary of the Invention The present invention provides a control algorithm where constant and optimal convergence is achieved by supplying an adaptation coefficient popt which varies in a calculated manner as a function of frequency and which results in an optimum, uniform, and constant adaptation or convergence rate. This method results in the minimum convergence time for the active noise or vibration control system.

In another aspect, the output from the adaptive LMS algorithm is gain limited according to the calculated relationship:

Gain limiting the output from the processor ensures the stability of the active control system by ensuring the output from the LMS controller will not be saturated, i.e., the output does not have, or will not cause, a square wave or near square wave shape which may impart harmonics or excite resonances in the dynamic system.Generally, one or more of the weights Wo or W1, preferably in an equal fashion, are reduced in magnitude in order to stabilize the value of gain G below a preselected threshold value of Gmsls It is a firther aspect of the invention to control the drift of the output device toward the saturation limits when singularities are encountered in the quadratic performance surface and other like conditions by applying leak factor r(k) as a function of f(k), where f(k) is a function of the weights W(k) or the gain G, where k represents a discreet time sample index.

Applying leakage will urge the output device away from the saturation limits.

It is a further aspect of the invention to control the drift of the output device toward saturation when singularities are encountered in the quadratic performance surface, and in combination, provide constant and optimal convergence by making the adaptation coefficient llopt vary in a calculated manner as a function of frequency. Thus, a constant adaptation or convergence rate is achieved and at the same time, a leakage factor r(k) which is a function of the filter weights W(k) or gain G is applied to urge the system away from the saturation limits by a increasing the leakage factor r(k) above a threshold preselected value of a.This combination results in no degradation to system performance while at normal operating conditions and implements leak only above a preselected threshold level a while at the same time ensuring optimal convergence rate by providing an optimum adaptation coefficient ,uopt for each frequency bin.

and, in combination, constant and optimal convergence is achieved by making the adaptation coefficient opt vary in a calculated manner as a function of frequency. Thus, optimal adaptation is achieved while maintaining the system gain G below a preselected limit Gmax such that the system never exceeds the saturation limits and/or is never operated within the nonlinear region of the output device.

It is an advantage of the present invention that the adaptation or convergence rate of the LMS algorithm is constant irregardless of frequency of input, thus minimizing the convergence time by providing an optimal convergence coefficient ptopt.

It is another advantage of the present invention that the algorithm output is stabilized before the physical system has exceeded the saturation limits by limiting the gain G of the system. This ensures that square wave like shapes are not inputted to the output device which may excite system resonances/modes or other system harmonics.

It is a further advantage of the present invention that the gain G of the system is limited above a predetermined level such that the electrical current provided to a output device, such as in an active mount or structural actuator, is limited. This prevents overheating of the output device, thus enhancing reliability and durability.

It is a further advantage of the present invention that the leak is increasingly applied above a predetermined level such that the output device is urged away from the saturation limits. This method has the advantage that leak is only applied near the saturation limits. Thus, control is not compromised for normal operating conditions.

It is a further advantage that the gain G of the system is limited to avoid exceeding the saturation limits while providing an optimum and constant convergence rate as a function of frequency.

The above-mentioned and further novel aspects and advantages of the present invention will become apparent from the accompanying descriptions of the preferred embodiments and attached drawings.

Brief Description of the Drawings The accompanying drawings which form a part of the specification, illustrate several embodiments of the present invention. The drawings and description together, serve to fully explain the invention. In the drawings, Fig. 1 is a block diagram illustrating a typical feedforward active control system; Fig. 2 is a block diagram illustrating the classical least mean square (LMS) algorithm; Fig. 3 is a block diagram illustrating the Filtered-x least mean square algorithm (Filtered-x LMS); Fig. 3A is a diagram illustrating a typical adaptation rate of an LMS system as a function of frequency; Fig. 3B is a diagram illustrating the a filter output to a output device which is saturated, i.e., which is clipped and has a square wave or near square wave shape;; Fig. 4A is a diagram illustrating a quadratic performance surface with no one localized minima, i.e., a trough; Fig. 4B is a diagram illustrating a quadratic performance surface with a global minima; Fig. 5 is a block diagram graphically illustrating gain limitation by adjusting weights Wg and W1 in a SISO system; Fig. 6 is a diagram graphically illustrating methods of reducing Wg and W1 by gain limitation to make sure Gma, is not exceeded; Fig. 7 is a block diagram illustrating an active system which utilizes gain limitation to stabilize temperature T below Tmax; Fig. 8 graphically illustrates an optimum and constant adaptation rate as a function of frequency;; Fig. 9 is a block diagram graphically illustrating a weight update law with factors for limiting the gain, providing constant adaptation rate, and for compensating for singularities; and Fig. 10 graphically illustrates various methods of providing leak to the filter weights.

Detailed Desription of the Invention In the following, it should be understood that the present invention is useful for controlling acoustic vibrations in air as well as mechanical vibrations in gases, liquids, or solids. Fig. 5 illustrates an aspect of the present invention where the adaptive filter weights are calculated and gain limited in such a manner so as to provide a signal u(k) to the output device (not shown) that is not saturated, i.e., gain limited such that the output falls inside the saturation limits.The weights Wg' and W1' are used to update the adaptive filter. W1 is the first filter weight coefficient which is multiplied with the sine component from the waveform generator and Wo is the second filter weight coefficient which is multiplied with the cosine component from the waveform generator. The multiplication takes place in block 26' representing an adaptive filter. The output from the adaptive filter is given by: u(k) = WO cos(kT) + Wl sin(kT) The same equation may also be expressed as: u(k) = Gcos(cokT + ) where G is the amplitude or gain of the output u(k).The gain G is calculated according to the equation:

Reducing Wg and W1 when the gain G is greater than a preselected threshold value of Gmax ensures that the output to the adaptive filter never exceeds the saturation limits.

The gain limitation calculation takes place in the processor and is simulated in a gain computation block 27'. The weights Wg' and W1' are not released to the adaptive filter until the gain G meets the requirement of the equation: G < Gmax Exceeding Gmax will produce a recurring signal which is saturated to some extent. This is because most control systems include electronic devices such as Digital-to-Analog (D/A) converters and analog filters. This devices act as saturation blocks and can only output voltages within a certain range based inherently on the electronic device design. Exceeding the voltage output will clip the device output. Although this is crude form of gain limitation, it may impart an undesirable square wave shape to the output signal.This is undesirable as it may excite the harmonics, modes, or other system resonances. Because of its shape, the signal will act as an impulse function, somewhat like a hammer tap exerted on a dynamic system. The impulse may excite multiple resonances or harmonics in the system as the impulse function contains broad frequency content and may contain enough power at any one frequency to elicit a system response.

The present invention limits the gain G before the signal enters the saturation block. The value of Gmax would be predetermined by system testing and would represent a value just below the saturation point of the system output. At least one of the weights Wo and W1 are reduced until G is less than or equal to Gmax. In this manner, the output is never saturated, because with proper selection of Gmax, the output device will never be sent a square wave shape or operate within the nonlinear range of the output device. Preferably both Wg and W1 are reduced equally.

The gain limitation method is used in a system for reducing the magnitude of a disturbance at a selected location or locations. Generally, the localized disturbance has been generated by a source of disturbance, i.e., a rotating unbalance, exhaust noise, propeller wash, or by any other noise or vibration sources. A signal is extracted from the source that is indicative of the characteristics of the source. This may include frequency, amplitude or phase information, as may required for the processor or waveform generator.

For a finite impulse response (FIR) implementation, a series of waves are input into an adaptive filter as supplied by a waveform generator, or if a digital implementation is used, then as supplied by the processor.

Either the processor or the waveform generator converts the signal indicative into a series of waves. These series of waves are generally a sine and cosine waves which are synchronized with the source of disturbance.

The series of waves are then inputted to an adaptive filter.

The first and second filter weight coefficients W1 and Wo are calculated according to a gradient descent method such as an LMS algorithm, Filtered-x LMS algorithm, or the like. The filter weights W1 and Wo are generally responsive to at least one error signal and/or input signal.

The output gain G is then calculated according to the abovementioned equation. If the gain exceeds Gmax, then one or more of the filter coefficients Wo and W1 are reduced until G is less than or equal to Gmax.

This results in an output from the processor, and thus, out of the adaptive filter which is gain limited such that the output does not exceed the saturation limits. In essence, the output cannot have a clipped form because of the implementation of gain limitation.

The filter weight coefficients Wg and W1 are applied, and operate on the series of waves, and as a result, generate a canceling signal u(k) which is supplied to a output device. The output device could be at least one of an electromagnetic actuator, piezoelectric actuator, hydraulic actuator, and a loudspeaker or other like device for providing a canceling disturbance at the at least one selected location. The result is that the active control system outputs can never extend beyond the prescribed saturation limits. It should be understood that infinite impulse response (IIR) implementations could also utilize the gain limitation method. However, the gain equations would be more complicated.

Fig. 6 graphically illustrates gain situations that are acceptable and unacceptable for the gain limitation method. The maximum allowable gain G is represented by a circle having a radius equal to Gmax within the W1, Wo plane. Location A represents a situation where the values of filter weights W1 and Wo will result in a value of G which is less than Gmax.

Thus, neither W1 or Wg would need to be adjusted and the weights would be passed to the filter without change. Location B represents a situation where the values of W1 and Wo will result in a value of G which is greater than Gmar. Therefore, one of weights W1 and Wo need to be reduced to bring G to or within the circle defined by Gmax. Reducing W1 and Who in equal proportions will result in a gain G at location C. Reducing W1 more than Wo will result in a gain G at location at D. Similarly, reducing Wg more than W1 will result in a gain G at location at E. Location C represents the new optimal solution that is stable and gain limited.This solution C is preferred because it contains the same phase angle as the solution B which is outside the Gmax circle.

Fig. 7 illustrates an active control system 20' which is similar to the system shown in Fig. 1 except that it incorporates gain limitation in order to ensure that a predetermined temperature Tmax of the output device is not exceeded. One of the problems associated with active control systems is that the actuators/mounts/output devices may get hot under certain conditions, such as when extreme vibrational environments are encountered wherein the actuators are operating at or near peak stroke conditions for extended periods of time. Heat is internally developed within the mounts/actuators due to resistive heating of the coils/elements within the active mounts/actuators and the like.

It is an aspect of the present invention to reduce the gain G below Gmax) when a condition is encountered where the instantaneous temperature T exceeds a predetermined threshold temperature value Tmax.

This allows the system to avoid an overheating condition which likely will affect system life, reliability, and linearity. As illustrated in Fig. 7, the temperature information is provided to the processor 34 by sensor 36 such as a thermister or the like.

As mentioned above, in classical LMS or Filtered-x LMS algorithms, the adaptation rate p is frequency dependent, and thus, the convergence rate can never be optimal. This aspect of the present invention overcomes this previous limitation and provides uniform, constant, and optimum convergence rate at every frequency bin.

In any practical controller, the frequency range that can be implemented is usually limited. The working frequency range is divided into a consecutive number K of frequency bands or intervals known as frequency bins. The center of each frequency bin is used to evaluate the plant transfer function P for the entire band, thus, providing an estimate of the transfer function for that bin. The real and imaginary coefficients for the real and imaginary matrices REi] and IM[ij for each frequency bin are obtained by a system identification (ID) process. This process is usually a sine sweep of frequency with a known sinusoidal input and the response thereto being recorded.The resulting plant transfer function can be stated as: P(Wj) = RE[i] + jIM[i] for i=1,2,3,...,K where PO'oi) = Plant transfer function for each frequency bin For single input, single output (SISO) systems, the RE[ij and IMFi] are vectors. For multiple input, multiple output (MIMO) systems, RE[i] and IM[i] are matrices having the N by M dimensions. Where N is the number of output devices and M is the number of sensors. The equation describing the plant transfer function for a MIMO system is as follows: P(IWj) = RE[i] + jIM[i] for

a=l,...,N s=l,...M output devices sensors frequency intervals where a = Output device index s = Sensor index N = Number of output devices M = Number of sensors All of the following methods will preserve the estimated phase angles of the plant transfer functions and only the magnitudes will be adjusted to achieve the optimum convergence.The following update equation was generally developed by Elliott, et al. for the Filtered-x LMS algorithm in a paper by Elliott, S.J., Stothers, I.M., and Nelson, P.A., 1987, entitled "A Multiple Error LMS Algorithm and Its Application to the Active Control of Sound and Vibration," IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. ASSP-35, No. 10, pp. 1423, 1434.

for Note that the Elliott et al. implementation includes a single adaptation coefficient KL for each output device, sensor, and frequency.

In a key aspect of the present invention, rate selection schemes are implemented to provide frequency dependent adaptation coefficients llopt to provide constant, uniform, and optimum convergence rate as a function of frequency.

Before the rate selection schemes are performed, however, it is desirable, but optional, to prefilter the MIMO REEji and IM(iJ matrices according to one of a TFGmin comparison method and a reconstructed G* matrix method.

The TFGmin comparison method is accomplished by choosing a minimum threshold transfer function gain TFGmin below which all values in the gain matrix G[iJ are set to zero. Then: IF ((RE +IM2 ) < (TGFmin2))THEN(RE = 0, and IM, = 0)

a=1,...,N #s=1,...,M Prefiltering will increase the numerical robustness by eliminating all the small numbers in the matrix which would result in large errors within the rate selection scheme. Furthermore, filtering out the small magnitude coefficients is desirable as they represent very "weak" coupling between the sensor-output device pairs and will only marginally affect the results.

To further reduce the computational burden, it is desirable, but optional, to further process the RE[il and IM[iJ matrices. This is accomplished by utilizing a reconstructed G* method whereby the Gain matrix is calculated and examined for each frequency bin according to the relationship:

where: Gas = Gain for each output device/sensor interaction REas = Real matrix for each output device/sensor interaction IMas = Imaginary matrix for each output device/sensor interaction Each row in the G(i] matrix represents the transfer function magnitudes associated with a given output device and each column in the G[iJ matrix represents the magnitudes associated with a given sensor.

First, looking at one row of the G[i] matrix at a time, the maximum or largest value is input into its corresponding place in a new matrix G*(iJ.

The rest of the values in that row are then set to zero in the Gain matrix.

Next, one column at a time, only the maximum or largest values from each column are placed in their corresponding places in the G*EiJ matrix and the rest of the values in that column are set to zero within the G*Eil matrix.

The result is a matrix G*[i] which contains only the maximum values occurring within the G[i] matrix. Finally, if the value of G*[i] at any position is zero, then the corresponding RE(ij and IM(i] coefficients of the Relic and IM[i] matrices are made to equal zero also according to the relation: IF{Gas = O} THEN (REas =0, and IMas =0}

s=l,...,M It should be noted that terms that are non-zero may be made to equal zero by the two prefiltering methods described, i.e., the TFGmin comparison method and reconstructed G* method.The key advantage of the reconstructed G* method of prefiltering is that the dominant coupling need not be known in advance. Furthermore, this method automatically handles the case where the dominant coupling changes with frequency.

Once these optional prefiltering steps have taken place the rate equalization scheme or gain limitation scheme may be implemented to arrive at a constant, uniform, and optimum convergence rate and/or gain limitation. Rate equalization is useful for use with a method of controlling a disturbance at at least one selected location which is generated by a source of disturbance.

Typical systems extract a signal indicative of the characteristics of said source of disturbance, convert that signal into a series of analog waves, or approximates of analog waves, and input the waves to an adaptive filter.

Within the processor, the optimum convergence coefficient opt is calculated. First, the maximum adaptation coefficient climax is calculated according to the relation:

Next, the optimal adaptation coefficient llopt is calculated according to the relationship: opt(#) = max(#).ss(N) where

and 2# #o= N where P (&commat;o) = scale factor o= = circular frequency corresponding to the disturbance N = &num; of samples/ period KLopt is calculated by the processor and represents an optimum adaptation coefficient which will provide optimum convergence for the active control system for each frequency bin.Furthermore, it should be recognized from the above equations that the optimum convergence coefficient is frequency dependent. As the plant transfer function P varies in frequency, so should uopt in order to provide optimal convergence at each frequency bin.

In order to accommodate the frequency dependent opt within the update equation, the function llopt may be incorporated into, blended with, or absorbed into the R matrix within the MIMO system. The new weight update equations become:

s=l,...,M where ROkas = REasCk +IMasSk R1kas=REasSk + ÎMas=Ck

########## Note that the coefficient Wopt has been absorbed into the real and imaginary RE and IM terms to derive the modified real and imaginary coefficients REasand ÎMas. The as subscripts denote output device/sensor pairs. The modified real and imaginary coefficients REas and IMas are used to calculate the modified response vectors ROkas and R1kas.The summation expressed above includes a measure of the gradient of the quadratic error surface. The weight update equation may be expressed in general terms as: W(k +1) = W(k) + ,uOp,V, where W(k) = adaptive filter weight coefficient opt = optimum adaptation coefficient Vk = gradient estimate of the quadratic performance surface Once the weights are calculated, they are inputted into an adaptive filter and the filter supplies a canceling signal to a output device for controlling the disturbance. It should be noted, that this convergence rate equalization method causes an adaptation rate for the active control system that is constant with respect to frequency as illustrated by Fig. 8.

There are two methods of absorbing the frequency dependent coefficients popt into the real and imaginary matrices RE[i] and IM[i], thus deriving or resulting in the new real and imaginary coefficients, RE and IM . This is achieved by calculating according to one of an equalization and a normalization process.

In the equalization process, each of the real REas and imaginary IMas coefficient pairs are scaled by a positive constant according to the relation:

a = s=1,,..,M where Ja = Number of non-zero transfer function for a given output device at a given frequency Non-zero optionally refers to coefficients of the transfer functions whose magnitudes are less than TFGmin such as when the aforementioned TFGmin comparison pre-filtering step is utilized or when the reconstructed G* method is used. Since both coefficients REas and IMas are scaled by the same constant, the phase information is preserved. Equalization here refers to equalizing the effect of each sensor on a given output device.The equalization method is best suited for systems with a single output device and multiple sensors.

The second method of rate equalization is best described as normalization. The normalization process will provide rate equalization but it will not attempt to equalize the effect of each sensor on a given output device. In this method, the coefficients are normalized with respect to the maximum gain of all the sensor transfer functions for a given output device, according to the relation:

la where

Both of these methods offer key advantages in that the optimum convergence rate is automatically selected.

In this aspect of the present invention, a different adaptation coefficient ,uopt must be used for each frequency bin. Thus, by utilizing the above scheme, as the plant transfer function varies as a function of frequency, so does the optimum adaptation coefficient llopt. This insures that the adaptation rate is constant for all frequencies. In one implementation, the convergence rates are equalized on a per-period basis.

However, they could be equalized on a time constant basis, as well.

Fig. 9 illustrates a block diagram of a system illustrating the update equation which allows for adjusting the revised filter weights W(k+1) according to the relationship: W(k+i)='P( r(Wk)+ ju Vk)) where v = Function allowing Gain Limitation r = Function which compensates for Singularities # = Function allowing Adaptation Rate optimization Vk = Estimate of the gradient This update law is useful for implementing the previously mentioned gain limitation, adaptation rate equalization, and compensating for singularities (to be described later) in various combinations.

Combinations of only < D and , only cP and r, are desired in most systems and may be implemented.

It should be understood that in most practical systems, it is desirable to provide both rate equalization and gain limitation of some form.

For example, a combination of gain limitation and rate equalization would be implemented according to the relation: Wk+1 = #{ (Wk) + #(Vk)}.

Further, implementations of rate equalization and singularity compensation are possible as well, and would be implemented according to the relation: w,*+, = ( (Wi)+!t (Vi) Generally gain limitation and singularity compensation would not be implemented separately as both methods are ultimately used to adjust the gain. Typically, one or the other would be used, but not both.

When a singularity is encountered, as mentioned above, the output device signal may drift toward the saturation limits. To solve this problem, a leakage factor r(k) is applied to the weights W(k) as a function of f(k) such that when the function f(k) approaches a predetermined functional value fmax, then the leakage factor r(k) is increased according to an update equation. By increasing the leakage factor r(k) asta function of f(k), the output device is maintained within operating limits, i.e., a saturation condition cannot occur. The function f(k) may be the gain G or the weight W(k). The key being that the leakage factor r(k) is adjusted according to the function f(k) such that the leakage is applied before the signal is inputted to the adaptive filter.The weight update equation used is: W(k t- 1) = (1- r(k)) W(k)- V(k) where W(k+1) = revised filter weights W(k) = Filter weight from a previous time step r(k) = leakage factor I1 = adaptation coefficient V(k)= gradient estimate It is notable that this method only applies leak when a condition is present where the output device may be exposed to a clipped signal or operated in a nonlinear portion of its stroke. Therefore, the system is not exposed to leak when normal operating conditions are present. As a result, performance is not compromised. The leakage factor r(k) may be implemented in a multitude of fashions, such as step-wise, linear, and non-linear. Fig. 10 graphically depicts the various leakage options.One possible method which is depicted by reference letter F of Fig. 10 is to let the leakage be updated in a stepwise fashion: r(k)= r@@ when 0 < f (k) # α and r(k) = rhi when α < f(k) # fmax.

where a = a predetermined value of f(k) between zero and fmax final = a maximum value of the function Another possible method which is depicted by reference letter G of Fig. 10 is to allow the leakage factor to be updated according to the linear relationship: r(k)= r@@ when 0 < f(k)# α and r(k) = ss2(f(k) - a) + r@@ when a < f(k) # fmax where

and P2 = an adjustment parameter The parameter a may be set near where the output device begin to operate in a non-linear fashion. The adjustment parameter can be used to adjust the slope or curvature and can be chosen such that r(k) reaches a maximum value at or near the saturation limit fmax. Care must be taken when choosing ss to maintain stability of the system. For example, care must be taken to insure that the function 1-pr(k) remains positive.

Yet another method which is depicted by reference letter H of Fig. 10 is to allow the output device weight to vary as a square function according to the relationship: r(k)= r@@ when 0 < f(k) #α and r(k) = ,B3 f(k)2 - (a)2 } + r,O when a < f(k) < fmax where ss3 = an adjustment parameter similar to P2 Another method of applying the leakage factor which is depicted by reference letter J of Fig. 10 is by allowing the leakage to be adjusted by a higher order function:: r(k)= r@@ when O < f(k) #α and r(k) = P,(f(k) - a) + rb when a < f(k) < fmix where p = an adjustment parameter similar to ss2 Still another method of controlling the leakage which is depicted by reference letter K of Fig. 10 where the leakage is adjusted by an exponential function according to the relation:: r(k) = rb when 0 < f(k) < a and r(k)=P, exp(f(k) - a) when a < f(k) < fx where > = an adjustment parameter similar to P2 In summary, in one aspect, the present invention provides a method for limiting the gain G the output to the adaptive filter when a gradient descent algorithm is utilized, such that the output signal is maintained.

within the saturation limits. It is a further aspect that optimum, uniform, and constant convergence rate can be achieved using an LMS or Filtered-x LMS algorithm. Further yet, singularities can be compensated for by providing appropriate leakage factor r(k) to the weight update equation such that the output device is maintained within the saturation limits. It is yet another aspect to gain limit the output to the adaptive filter while, in combination, providing constant and optimal convergence rate. This can be accomplished by applying leakage factor r(k) or limiting the value of G such that G is less than or equal to a predetermined threshold value Gmax, and in combination, providing an optimal adaptation coefficient )lopt for each frequency bin to provide optimal convergence rate as a function of frequency.

While several embodiments of the present invention have been described in detail, various modifications, alteratiors, changes and adaptations to the aforementioned may be made without departing from the spirit and scope of the present invention defined in the appended claims. It is intended that all such modifications, alterations, changes and adaptations be considered part of the present invention.

Claims

1. A method of controlling a disturbance at at least one selected location which has been generated by a source of disturbance, comprising the steps of: (a) extracting a reference signal indicative of characteristics of said source of disturbance; (b) converting said reference signal into a series of waves; (c) inputting said series of waves into an adaptive filter; (d) calculating a first filter weight W1 and a second filter weight Wo for said adaptive filter according to an algorithm such that said first filter weight W1 and said second filter weight Wo are responsive to at least one of said reference signal and at least one error signal;; (e) calculating an output gain G according to the equation

(f) reducing at least one of said first filter weight W1 and said second filter weight Wg if said output gain G > Gmax, where said Gmax is a preselected threshold gain value, such that G < (g) inputting said first filter weight W1 and said second filter weight Wg to said adaptive filter to operate on said series of waves, thereby generating a canceling signal; and (h) outputting said canceling signal to an output device for controlling said disturbance at said at least one selected location whereby said canceling signal is gain limited to fall within a saturation limit of said canceling signal.

2. A method of Claim 1 wherein numerical robustness of a RE[i] matrix and a IM[i] matrix is increased by: (a) calculating a gain

where REas is a real matrix and IMas is an imaginary matrix; (b) comparing Gas to a minimum threshold gain TFGmin; and IF {(REas2 + IMas2) < (TFGmin)}THEN{REas = 0, and IMas = O}

a=1 N s=I M

3. A method of Claim 1 wherein said adaptive filter is a finite impulse response (FIR) adaptive filter.

4. A method of Claim 1 wherein said algorithm is one of an LMS and a Filtered-x LMS algorithm.

5. A method of Claim 1 wherein said gain G is adjusted to stabilize a temperature T below a maximum temperature value of Tom at

6. A method of Claim 1 wherein said output device is one of an electromagnetic actuator, piezoelectric actuator, hydraulic actuator, and a loudspeaker.

7. A method of controlling a disturbance at at least one selected location which is generated by a source of disturbance, comprising the steps of: (a) extracting a reference signal indicative of characteristics of said source of disturbance; (b) converting said reference signal to a series of waves; (c) inputting said series of waves into an adaptive filter; (d) calculating max(#), where max() is a maximum adaptation coefficient; (e) calculating

where ss(o) is a scale factor, and '#o is 2# divided by a number of samples per period N; (f) calculating an optimum adaptation coefficient ilopt for each frequency bin according to the equation opt & ommat;;= max.ss(#0) where ,Umax is a function of a RE[i) matrix and a IM[ii matrix; (g) calculating revised filter weights W(k+1) based upon the equation W(k + 1) = W(k)+ lop,V where W(k) are filter weights from a previous time step and Vk is a measure of a gradient of a quadratic error surface; (h) inputting said revised filter weights to said adaptive filter; (i) outputting a canceling signal to an output device for controlling said disturbance thereby causing an adaptation rate which is constant with respect to frequency.

8. A method of controlling a disturbance of Claim 7 wherein Filtered-x LMS control is used and

where max(#) is a maximum adaptation coefficient, P(w) is a plant transfer function, and A is the magnitude of an input xk.

9. A method of controlling a disturbance of Claim 7 wherein LMS control is used and

where ,Umax() is a maximum adaptation coefficient, and A is the magnitude of an input xk.

10. A method of controlling a disturbance of Claim 7 including the further steps of: (a) calculating a magnitude of a gain G[i] matrix for each said frequency bin; (b) selecting a largest value of gain G from each row of said G[i] matrix and a largest value from each column of said G[i] matrix and placing said largest values in a G* matrix, placing zeros in said G* matrix corresponding to locations of all smaller values in said rows and said columns of said G* matrix; and (c) setting all real components of said RE[i] matrix and all imaginary components of said IM[i] matrix to zero which correspond to said zero values of said G* matrix, for each said frequency bin whereby only dominant terms of said RE[ij and IM[i] matrices remain.

11. A method of controlling a disturbance of Claim 7 including the further steps of: (a) absorbing said optimum adaptation coefficient Adopt into said real matrix RE[i] and said imaginary matrix IM[i] to derive modified real and imaginary matrices Reasand IMas; (b) calculating said modified and imaginary matrices REasand ÎMas according to one of an equalization process and normalization process; (c) calculating modified response vectors to be used in a first filter weight calculation according to a relationship R1kas=REasSk-IMasCk for

#a=l,...,N s=1,...,M; and (d) calculating modified response vectors to be used in a second filter weight calculation according to a relationship ROkas = REasCk-IMasSk for

MN

12. A method of controlling a disturbance of Claim 11 wherein said equalization process is calculated according to a relation

a =1 N s=1,...,M, where Ja is a number of non-zero transfer functions for output device a at a given frequency.

13. A method of controlling a disturbance of Claim 11 wherein said normalization process is calculated according to a relationship

1S where

a=1,...,N and where Ja is a number of non-zero transfer functions for output device a at a given frequency.

14. A method of controlling a disturbance at at least one selected location which has been generated by a source of disturbance, comprising the steps of: (a) extracting a reference signal indicative of characteristics of said source of disturbance; (b) converting said reference signal into a series of waves; (c) inputting said series of waves into an adaptive filter; (d) calculating revised filter weights W(k+1) based upon the equation W(k +1) = (1- r(k)} W(k) + V(k), where Vk is a measure of a gradient of a quadratic error surface, r(k) is a leak factor, W (k) are weights from a previous time step, and is an adaptation coefficient;; (e) inputting said revised filter weights W(k+l) into said adaptive filter to generate a canceling signal; (f) outputting said canceling signal to a output device for controlling said disturbance at said selected location; and (g) controlling said revised filter weights W(k+1) by applying leakage factor r(k) to said weights W(k) such that said leakage factor r(k) increases as a function f(k) approaches a predetermined threshold value fmax whereby said leak factor r(k) increases as a function of said function f(k) to prevent said canceling signal from becoming saturated.

15. A method of controlling a disturbance of Claim 14 wherein f(k) is one of a gain value G(k) and a weight value W(k).

16. A method of controlling a disturbance of Claim 16 wherein said output device weight r(k) = rtO when 0 < f(k) < a and r(k) = r when a < f(k) < flax, where r10 is a low value of said leakage factor, rhi is a high value of said leakage factor, fmax is a maximum value of function f(k), and a is a preselected intermediate value of function f(k).

17. A method of controlling a disturbance of Claim 16 wherein said output device weight r(k) = r@@ when 0 < f (k) #α and r(k)=ss2{f(k)-α}+rlo when a < f(k) < fmax, where r10 is a low value of said leakage factor, fmax is a maximum value of function f(k), a is a preselected intermediate value of function f(k), and P2 is an adjustment parameter.

18. A method of controlling a disturbance of Claim 16 wherein said output device weight r(k)=r" when 0 < f(k)# α and r(k) = ss3{ f(k)2 - (a)2 )+ r, when a < f(k) < fmax, where r10 is a low value of said leakage factor, fmax is a maximum value of function f(k), a is a preselected intermediate value of function f(k), and P3 is an adjustment parameter.

19. A method of controlling a disturbance of Claim 16 wherein said output device weight r(k) = rlo when 0 < f(k)# α and r(k) = P4 (f(k) - razz + rio when a < f(k) < f1flax, where r1O is a low value of said leakage factor, fmax is a maximum value of function f(k), a is a preselected intermediate value of f(k), and P4 is an adjustment parameter.

20. A method of controlling a disturbance of Claim 16 wherein said output device weight r(k)= r@@ when 0 < f(k)# α and r(k) = Ps exp(f(k) - a) when a < f(k) < frnax, where r1O is a low value of said leakage factor, fmax is a maximum value of function f(k), a is a preselected intermediate value of f(k), and ss5 is an adjustment parameter.

21. A method of controlling a disturbance at at least one selected location which is generated by a source of disturbance, comprising the steps of: (a) extracting a reference signal indicative of characteristics of said source of disturbance; (b) converting said reference signal to a series of waves; (c) inputting said series of waves into an adaptive filter; (d) calculating m a x (#) according to one of

when Filtered-x LMS control is utilized and

when classical LMS control is utilized, where ,amaxf ) is a maximum adaptation coefficient, Pf) is a plant transfer function, and A is the magnitude of an input xk;; (e) calculating

where P(a)o) is a scale factor, and a)o is 2# divided by the number of samples per period N; (f) calculating an optimum adaptation coefficient opt for each frequency bin according to the equation opt(#)= max.ss(#o) where max is a function of a real matrix RE[i] and a imaginary matrix IM[i]; (g) absorbing said optimum adaptation coefficient adopt into said real matrix RE[ii and said imaginary matrix IM[i] to derive modified real and imaginary coefficients Remand as;; (h) calculating said RE and 1M according to one of

a=l,...,N s=l,...,M, where Ja is a number of non-zero transfer functions for output device a at a given frequency, and

########## where Gas =max{REas+IMas} a=1,...,N and where Ja is a number of s=1.M non-zero transfer functions for output device a at a given frequency;; (i) calculating a modified response vector to be used in a first filter weight calculation according to a relationship R1kas =REasSk-IMasCk for

a=l,..., N #s =1,...,M, (j) calculating a modified response vector to be used in a second filter weight calculation according to a relationship ROkas = REasCk-ÎMasSk for

a=1,...,N #s=1,...,M, (k) calculating said first and said second filter weights based upon the equations

l.s=l,...,M where E. is at least one error signal value; (1) inputting said first and said second filter weights to said adaptive filter; and (m) outputting a canceling signal to an output device for controlling said disturbance; thereby causing an adaptation rate which is constant with respect to frequency.

22. A method of controlling a disturbance substantially as hereinbefore described, with reference to the accompanying drawings.